Eseandre Mordi
⬤A new cross-country benchmark has mapped where leading AI models stand politically. The analysis tested GPT-5, Gemini 2.5 Pro, Claude 4.5 Sonnet, Grok-4, Kimi K2 Thinking and Magistral Medium by having them "vote" in real elections across eight nations. Results show nearly all systems landing in the libertarian-left quadrant, revealing a consistent ideological tilt instead of true neutrality.
⬤The models were told to act like local voters and their choices were compared against actual election results. Despite differences in culture and politics across countries, the AI systems kept leaning toward center-left, progressive candidates. Gemini 2.5 Pro and GPT-5 sit near the center of the libertarian-left zone, while Claude and Kimi position slightly further left. Magistral Medium clusters with them, and Grok-4 is the only model drifting somewhat toward libertarian-right territory, though it still stays centrist on economic issues.
⬤The study highlights a clear gap between AI preferences and real-world voting. Even when models roleplay as residents of specific countries, their "votes" don't match actual voter behavior. This raises questions for Google, OpenAI, Anthropic, xAI and Mistral about training data, alignment methods and whether model development accidentally shapes political reasoning. The fact that different brands cluster together suggests shared biases built in during training.
⬤These findings matter because ideological leanings in AI can affect user trust and regulatory debates. As AI gets embedded in search engines and productivity tools, identifiable political tendencies may impact transparency standards across the tech sector. The benchmark adds fresh context to discussions about measuring and monitoring political behavior in AI systems.
Eseandre Mordi
